Thuật ngữ AI
Từ điển đầy đủ về Trí tuệ nhân tạo
Fair prediction
Statistical paradigm combining conditional parity and calibration to ensure fair predictions between groups. Fair prediction aims to balance inherent trade-offs between different fairness criteria in predictive models.
Prediction calibration
Statistical property ensuring that prediction scores accurately reflect true probabilities for all groups. Perfect calibration ensures that a 70% score corresponds to exactly 70% positive outcomes, regardless of the group considered.
Algorithmic counterfactual fairness
Fairness approach examining decisions that would have been made if protected characteristics were different. Counterfactual fairness evaluates whether similar individuals would receive equivalent outcomes by changing only their demographic attributes.
False positive rate parity
Fairness criterion requiring all demographic groups to have equivalent type I error rates. This metric ensures that no group systematically experiences more false accusations or unjustified rejections.
Individual fairness
Ethical principle stating that similar individuals should receive similar treatments, regardless of their group membership. Individual fairness contrasts with group fairness by focusing on specific cases rather than statistical aggregates.
Bias mitigation
Set of mathematical techniques aimed at correcting systematic disparities in data and predictive models. Mitigation includes pre-processing, in-processing, and post-processing methods to achieve algorithmic fairness.
Indirect discrimination
Form of algorithmic discrimination where proxy variables illegitimately substitute explicitly excluded protected characteristics. Indirect discrimination emerges when statistical correlations reproduce inequalities without directly using sensitive attributes.